Back to Blog

How I Built a Blueprint Estimator Using Claude AI

A South Florida general contractor was spending 3–4 hours per bid doing manual takeoffs from blueprint PDFs. I built an AI tool that does the same job in under 5 minutes. Here's exactly what I built and how it works.

---

The problem

In construction, "takeoff" is the process of reading a blueprint and manually cataloging everything that needs to be purchased or built — lumber, concrete, electrical, framing, hardware. It's tedious, detail-oriented work, and it has to be done for every single bid.

For a busy GC, this means 3–4 hours per project just to generate a starting estimate — before any pricing, before any subcontractor calls, before anything else. On a week with three active bids, that's nearly a full day gone to data extraction.

The manual process also introduces errors. Misread dimensions. Missed line items. Math done by hand. Every mistake either costs money on the job or costs the bid entirely.

The solution

I built a tool using Claude AI (Anthropic's API) and Python that accepts a blueprint PDF as input and outputs a structured Excel bid sheet — populated with material categories, quantities, unit descriptions, and labor line items — in under 5 minutes.

The workflow looks like this:

1. The contractor uploads a blueprint PDF through a simple browser interface.
2. The tool sends the PDF to Claude AI, which reads the document and extracts structured data — room dimensions, material callouts, spec notes, trade breakdowns.
3. Claude returns a structured JSON object with all the line items.
4. A Python script (using openpyxl) writes that data into a formatted Excel bid sheet — one tab for materials, one for labor, with totals and markup formulas already built in.
5. The contractor downloads the completed sheet and starts pricing from a clean baseline.

What Claude AI actually does here

This is the part that surprises most people. Claude isn't just doing keyword extraction — it's reading the blueprint the way a human estimator would. It understands spatial relationships, cross-references spec callouts with floor plan annotations, and applies construction domain knowledge to fill in implied line items that aren't explicitly labeled.

For example, if a blueprint shows a bathroom addition with tile callouts, Claude infers the need for cement board substrate, tile adhesive, and grout — even if those aren't spelled out on the drawing. A keyword scraper would miss those entirely.

The stack

• Claude API (Anthropic) — document reading and structured data extraction
• Python + FastAPI — backend server handling file uploads and API calls
• openpyxl — writing the formatted Excel output
• Simple HTML/JS frontend — drag-and-drop upload interface running locally
• launchd — keeps the server running automatically on the Mac mini

The whole thing runs locally on a Mac mini — no cloud subscription, no data leaving the building, no monthly SaaS fee beyond the Claude API usage (which runs about $2–$8 per blueprint processed, depending on size).

The result

What used to take 3–4 hours now takes 5–10 minutes, including review time. The contractor still checks the output — and occasionally adds line items the AI missed — but they're working from a 90%+ complete baseline instead of a blank sheet. The error rate dropped significantly because the math is automated and the line items are consistent.

On a week with three active bids, that's roughly 10 hours recovered. At the contractor's effective hourly rate, that's meaningful money — and it means more bids can be turned around faster, which means more work won.

Could this work for your business?

If you have any repeating process that involves reading a document and transferring information somewhere else — the answer is almost certainly yes. The blueprint estimator is one flavor of a pattern I've built variations of for invoices, contracts, job applications, and medical intake forms.

The specifics change. The approach doesn't.